Why

The two PRs below this one establish the composition behavior for explicit unmapped boundaries and coarse same-line mappings. The hot loop only needs numeric token fields, but it still constructs source-view wrappers for every final and traced token.

This PR iterates raw Tokens instead. It keeps strict lookup as the common path and reads the first token on the same line only when strict lookup misses. The behavior and output are unchanged; the tests added by the two lower PRs cover both exceptional paths.

Benchmark

Command:

cargo bench -p bench --bench bench -- sourcemap/collapse_codegen_chain \
  --warm-up-time 2 --measurement-time 5 --sample-size 50

Environment: Apple M3 Pro, Darwin 25.5.0 arm64, rustc 1.97.0.

The repository's 52,004-token, three-map composition benchmark measured:

source-view tokens:  [2.2210 ms 2.2308 ms 2.2416 ms]
raw tokens, run 1:   [2.0215 ms 2.0285 ms 2.0368 ms]  (-8.93% median)
raw tokens, run 2:   [2.0089 ms 2.0179 ms 2.0298 ms]  (-9.28% median)

Both comparisons were statistically significant (p = 0.00). These numbers describe this isolated benchmark, not an end-to-end build speedup.

Validation

• cargo test -p rolldown_sourcemap
• just test-rust
• both plugin regression fixtures from the lower PRs
• cargo clippy -p rolldown_sourcemap --all-targets -- --deny warnings